Fig. 5.

Striatal and FPC activity predicts Bayesian updating performance. a Displays the effect of contrasting behavioural and neural effects across participants forming a contrast of contrasts. The lower the difference between neural short-term (RPE) and long-term (D _KL) representations in bilateral caudate (right 9, 2, 10 mm, z-score = 4.43; left −11, 3, 13 mm, z-score = 4.41) and FPC (right 17, 66, 5 mm, z-score = 4.34; left −24, 64, 9 mm, z-score = 4.59), the more participants updated their beliefs in accordance with predictions from the Bayesian, rather than the RL model. Conjunction analyses b with the main effects of RPE (red) and D _KL (blue) showed overlap mainly within regions that were primarily activated by Bayesian model update, although medial striatum and FPC activity additionally overlapped with the main effect of RPE. This was confirmed by a three-way conjunction analysis which revealed that indeed the medial striatum, as well as a small region in FPC was significantly activated by model-free and inference-based learning and reflective of the degree of influence both models exerted over participants’ belief updates when compared on group level (c, yellow). Colour bars indicate z-scores, plots are cluster corrected at p < 0.05